Disk Types and Performance

There are the following types of EVS disks, each with different levels of I/O performance: Extreme SSD, General Purpose SSD V2, Ultra-high I/O, General Purpose SSD, High I/O, and Common I/O. EVS disks differ in performance and price. You can choose whichever disk type that is the best fit for your applications.

Extreme SSD EVS disks use the congestion control algorithms for Remote Direct Memory Access (RDMA) deployments. An extreme SSD disk can reach up to 1,000 MiB/s of throughput and with extremely low single-channel latency.

EVS Performance

EVS performance metrics include:

IOPS: number of read/write operations performed by an EVS disk per second
Throughput: amount of data read from and written into an EVS disk per second
Read/write I/O latency: minimum interval between two consecutive read/write operations on an EVS disk

**Table 1** EVS performance data
Parameter	Extreme SSD	General Purpose SSD V2	Ultra-high I/O	General Purpose SSD	High I/O	Common I/O (Previous Generation Product)
Max. capacity (GiB)	System disk: 1,024 Data disk: 32,768	System disk: 1,024 Data disk: 32,768	System disk: 1,024 Data disk: 32,768	System disk: 1,024 Data disk: 32,768	System disk: 1,024 Data disk: 32,768	System disk: 1,024 Data disk: 32,768
Short description	Superfast disks for workloads demanding ultra-high bandwidth and ultra-low latency	SSD-backed disks allowing for tailored IOPS and throughput and targeting for transactional workloads that demand high performance and low latency	High performance disks excellent for enterprise mission-critical services as well as workloads demanding high throughput and low latency	Cost-effective disks designed for enterprise applications with medium performance requirements	Disks suitable for commonly accessed workloads^f	Disks suitable for less commonly accessed workloads
Typical workloads	Database workloads Oracle SQL Server ClickHouse AI workloads	Enterprise OA and virtual desktops Large-scale development and test environments Transcoding services System disks Medium- and large-sized databases (SQL Server, Oracle, NoSQL, and PostgreSQL)	Transcoding services I/O-intensive workloads NoSQL Oracle SQL Server PostgreSQL Latency-sensitive applications Redis Memcache	Enterprise OA Medium-scale development and test environments Small- and medium-sized databases Web applications System disks	Common development and test environments	Applications demanding large capacity, medium read/write speed, but having fewer transactions Common office applications Lightweight development and testing Not recommended to be used as system disks
Max. IOPS^a	128,000	128,000	50,000	20,000	5,000	2,200
Max. throughput^a (MiB/s)	1,000	1,000	350	250	150	50
Burst IOPS limit^a	64,000	N/A	16,000	8,000	5,000	2,200
Disk IOPS^c	Min. [128,000, 1,800 + 50 x Capacity (GiB)]	You preconfigure an IOPS ranging from 3,000 to 128,000. This IOPS must also be less than or equal to 500 times the capacity (GiB).	Min. [50,000, 1,800 + 50 x Capacity (GiB)]	Min. [20,000, 1,800 + 12 x Capacity (GiB)]	Min. [5,000, 1,800 + 8 x Capacity (GiB)]	Min. [2,200, 500 + 2 x Capacity (GiB)]
Disk throughput^b (MiB/s)	Min. [1,000, 120 + 0.5 × Capacity (GiB)]	You preconfigure a throughput ranging from 125 to 1,000. This throughput must also be less than or equal to the IOPS divided by 4.	Min. [350, 120 + 0.5 × Capacity (GiB)]	Min. [250, 100 + 0.5 × Capacity (GiB)]	Min. [150, 100 + 0.15 × Capacity (GiB)]	50
Single-queue access latency^d (ms)	Sub-millisecond	1	1	1	1–3	5–10
API Name^e	ESSD	GPSSD2	SSD	GPSSD	SAS	SATA

a: The maximum IOPS, maximum throughput, and burst IOPS limit all include both read and write operations. So, maximum IOPS = read IOPS + write IOPS.

b: Take ultra-high I/O for example: The baseline throughput is 120 MiB/s. The throughput increases by 0.5 MiB/s for every one GiB added until it reaches the maximum throughput 350 MiB/s.

c: Take ultra-high I/O for example: The baseline IOPS is 1,800. The IOPS increases by 50 for every one GiB added until it reaches the maximum IOPS 50,000.

d: A single queue indicates that the queue depth or concurrency is 1. The single-queue access latency is the I/O latency when all I/O requests are processed sequentially. The values in the table are calculated with 4 KiB data blocks.

e: This API name is the value of the volume_type parameter in the EVS API. It does not represent the type of the underlying hardware device.

f: High I/O disks (except for those created in dedicated storage pools) are HDD-backed disks. They are suitable for applications with commonly accessed workloads. The baseline throughput of a high I/O disk is 40 MiB/s per TiB, and the maximum throughput of a high I/O disk is 150 MiB/s. If your applications have high workloads, it is recommended that you choose SSD-backed disks which have higher specifications.

EVS disk performance is closely related with the data block size:

If data blocks are all the same size, a disk can achieve either the maximum IOPS or maximum throughput depending on which one is reached first.
If data blocks are of different sizes, the maximum performance metric that a disk can achieve varies:
- For small data blocks, such as 4 KiB or 8 KiB, a disk can reach the maximum IOPS.
- For data blocks greater than or equal to 16 KiB, a disk can reach the maximum throughput.

Table 2 uses an ultra-high I/O disk as an example. In theory, when the size of an ultra-high I/O disk is at least 964 GiB, the disk theoretically can reach either the maximum IOPS 50,000 or the maximum throughput 350 MiB/s. However, this is not the case in practice. The maximum IOPS and maximum throughput that a disk can reach also vary with the data block size.

**Table 2** Maximum performance of an ultra-high I/O EVS disk
Data Block Size (KiB)	Max. IOPS	Max. Throughput (MiB/s)
4	About 50,000	About 195
8	About 44,800	About 350
16	About 22,400	About 350
32	About 11,200	About 350

Disk IOPS Calculation Formula

Disk IOPS = Min. (Maximum IOPS, Baseline IOPS + IOPS per GiB x Capacity)

This formula does not apply to General Purpose SSD V2 disks.

For a General Purpose SSD V2 disk: You preconfigure an IOPS ranging from 3,000 to 128,000. This IOPS must also be less than or equal to 500 times the capacity (GiB).

Take an ultra-high I/O EVS disk with a maximum IOPS of 50,000 for example.

If the disk capacity is 100 GiB, the disk IOPS is calculated as follows: Disk IOPS = Min. (50,000, 1,800 + 50 × 100)

The disk IOPS is 6,800, the smaller of the two values (50,000 and 6,800).
If the disk capacity is 1,000 GiB, the disk IOPS is calculated as follows: Disk IOPS = Min. (50,000, 1,800 + 50 × 1,000)

The disk IOPS is 50,000, the smaller of the two values (50,000 and 51,800).

Disk Burst Capability and Principles

EVS disks have a burst capability. A small-capacity disk can surpass its official maximum IOPS for a short period of time. This IOPS applies to each disk individually.

Disks with burst capability are well-suited for speeding up server startup. In most cases, system disks are fairly small, so their basic IOPS is fairly low. For example, the IOPS of a 50-GiB ultra-high I/O disk without burst can only reach up to 4,300 IOPS (Min. (50,000, 1,800 + 50 x Capacity)). But with burst capability, its IOPS can burst up to 16,000.

The following example uses an ultra-high I/O EVS disk with the IOPS burst limit of 16,000.

If the disk capacity is 100 GiB, the disk has a maximum IOPS of 6,800, but it can burst to 16,000 IOPS in a certain duration.
If the disk capacity is 1,000 GiB, the disk has a maximum IOPS of 50,000. The disk maximum IOPS already exceeds its burst IOPS 16,000, and the disk does not use the burst capability.

The following describes the burst IOPS consumption and reservation.

A token bucket is used to handle burst I/O operations. The number of initial tokens in the bucket is calculated as follows:

Number of initial tokens = Burst duration x IOPS burst limit

In the following example, a 100-GiB ultra-high I/O EVS disk is used, and the fixed burst duration is 1800s. Therefore, the number of initial tokens is 28,800,000 (1,800 x 16,000).

Token production rate: This rate equals the disk maximum IOPS, which is 6,800 tokens/s.
Token consumption rate: This rate is based on the I/O usage. Each I/O request consumes a token. The maximum consumption rate is 16,000 tokens/s, which is the larger value of the disk burst IOPS and the maximum IOPS.

Consumption principles

When tokens are consumed faster than they are produced, the number of tokens decreases accordingly, and eventually the disk IOPS will be consistent with the token production rate (the maximum IOPS). In this example, the disk can burst for approximately 3,130 seconds [28,800,000/(16,000 - 6,800)].

Reservation principles

When tokens are consumed more slowly than they are produced, the number of tokens increases accordingly, and the disk regains burst capability. In this example, if the disk is suspended for approximately 4,235 seconds (28,800,000/6,800), the token bucket will be filled up with tokens.

As long as there are tokens in the token bucket, the disk has the burst capability.

Figure 1 shows the token consumption and reservation principles. The blue bars indicate the disk IOPS usage, the green dashed line represents the maximum IOPS, the red dashed line indicates the IOPS burst limit, and the black curve indicates the changes of the number of tokens.

As long as there are tokens, the disk IOPS can exceed 6,800 and can burst up to 16,000, the IOPS burst limit.
When there are no more tokens, the disk loses the burst capability, and the disk IOPS can reach up to 6,800.
Any time the disk IOPS is less than 6,800, the number of tokens starts to increase, and the disk regains the burst capability.

Figure 1 Burst capability diagram